Artificial intelligence-based modeling and optimization of heat-affected zone and magnetic property in pulsed laser cutting of thin nonoriented silicon steel

Abstract

Laser machining has been emerging as a powerful alternative for cutting thin metal substrates. In this study, the application of pulsed laser cutting of a thin nonoriented silicon steel, with a thickness of 0.1 mm, was studied. The four processing parameters considered were laser power, cutting speed, pulse repetition rate, and processing environment. The two outputs to be measured were the extent of heat-affected zone (HAZ) and deviation of magnetic flux density (MFD) from initial value. Each input was designed with three levels and the three processing environments were air, deionized water, and sodium chloride solution. Based on the experimental design of the L27 Taguchi method, 27 parameter sets out of the total of 81 sets were used for the experiment. Results show that HAZ and MFD were negatively correlated. Compared with processing in air, cutting in the liquid could effectively reduce the HAZ. In the 27 experimental cases, the achieved minimum HAZ was 34.5 μm that corresponded to retaining 99% of initial MFD. The importance of the input was analyzed by the random forest method. The most and second significant parameters were laser power and environmental condition and their importance levels were 50.82% and 40.99%, respectively. Four artificial intelligence (AI) prediction models, full quadratic multiple regression analysis, artificial neural network, random forest, and extreme learning machine (ELM), were established based on randomly selecting 80% of the 27 data sets for training and the remaining 20% for testing. Model verification was executed by arbitrarily taking 10 additional new predictive parameter sets, from the remaining 54 parameter sets, for experiments. After comparing the predicting and experimental results, ELM model was found to have the best forecast performance. Thus, it was chosen as the target model for output optimization by the genetic algorithm method (GA). Through implementing the predicted optimal processing parameters from the resulting ELM-GA algorithm for the confirmation experiment, the obtained MFD and HAZ were 1.639 T and 30.41 μm, respectively, which were very close to that of the predicted optimal outputs, 1.640 T for MFD and 29.90 μm for HAZ.